GPT-6 caching update cuts cached input costs by up to 90%

On September 22, OpenAI gave developers explicit cache breakpoints for GPT-6. A breakpoint marks where the reusable start of a prompt ends. The same update brings higher cache hit rates, new diagnostics and controls for latency and cost, OpenAI said in its announcement, and it cuts cached input token costs by up to 90%. Newsroom America also reports a caching dashboard.

OpenAI announced the caching update on the same day it halved the list prices of GPT-6 Sol and Luna. GPT-6 Sol costs $2/$10 per million input/output tokens, down from $4/$20 for GPT-5.6 Sol. So GPT-6 got cheaper in two places: the list price, and the part of each request that repeats. Agent loops and chat apps resend the same system prompt, tool definitions and history on every call, so for them the second cut probably covers more tokens than the first.

Caching itself isn’t new. An earlier version of OpenAI’s guide promised up to 90% off input token costs and up to 80% lower latency, with caching running automatically on prompts of 1,024 tokens or longer, on gpt-4o and newer models, with no code changes and no extra fees. OpenAI’s current prompt caching guide describes the cached-input rate as “discounted up to 95%”.

The guide also explains the mechanism. When a model processes input tokens, it computes intermediate key-value states. Caching saves those states for an unchanged prefix, so a later request with the same prefix can reuse them and only processes the new input. The cache holds those tensors, not the tokens themselves. According to the guide, reuse needs the entire rendered prefix to match, and that prefix includes OpenAI’s own instructions, developer messages, tool definitions and conversation history. Changing the model, tools, parallel_tool_calls, text.format, reasoning.effort, text.verbosity or the compaction setting context_management can prevent reuse from the first changed token onward.

OpenAI’s guide describes breakpoints the same way. The first request writes an eligible prefix to the cache. Later requests look for the longest matching cached prefix, working backward through eligible breakpoints until they find a match.

Billing follows a three-way split, per the guide: each input token is charged at the uncached, cached or cache-write rate. The guide says cache-write pricing is not an additive fee. Agents API model calls use the same caching behavior as the Responses API.

Keeping a session open doesn’t guarantee a hit, the guide warns, though reusing context within a session can preserve a shared prefix. To check, OpenAI points to a Prompt Caching Dashboard that tracks cache read hit rates and a Prompt Cache Diagnostics tool for working out why a request missed.